Fire detection system using random forest classification for image sequences of complex background
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Fire detection system using random
forest classification for image sequences
of complex background
Onecue Kim
Dong-Joong Kang
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Optical Engineering 52(6), 067202 (June 2013)
Fire detection system using random forest classification
for image sequences of complex background
Onecue Kim Abstract. We present a fire alarm system based on image processing that
Dong-Joong Kang detects fire accidents in various environments. To reduce false alarms that
Pusan National University frequently appeared in earlier systems, we combined image features
Department of Control and Automation including color, motion, and blinking information. We specifically define
Republic of Korea the color conditions of fires in hue, saturation and value, and RGB
E-mail: djkang@pusan.ac.kr color space. Fire features are represented as intensity variation, color
mean and variance, motion, and image differences. Moreover, blinking
fire features are modeled by using crossing patches. We propose an algo-
rithm that classifies patches into fire or nonfire areas by using random for-
est supervised learning. We design an embedded surveillance device
made with acrylonitrile butadiene styrene housing for stable fire detection
in outdoor environments. The experimental results show that our algorithm
works robustly in complex environments and is able to detect fires in real
time. © The Authors. Published by SPIE under a Creative Commons Attribution 3.0 Unported
License. Distribution or reproduction of this work in whole or in part requires full attribution of the
original publication, including its DOI. [DOI: 10.1117/1.OE.52.6.067202]
Subject terms: fire detection; random forest; visual surveillance; embedded camera.
Paper 130413 received Mar. 17, 2013; revised manuscript received May 3, 2013;
accepted for publication May 13, 2013; published online Jun. 18, 2013.
1 Introduction environments. In most of the previous studies, it is necessary
Conventional fire detection systems were designed to detect to determine whether an individual pixel is a fire or nonfire
smoke, heat, and radiant energy from a fire using infrared, pixel. However, red objects that have a hue as fire are often
optical, and ion sensors.1–3 However, these methods have mistaken for fire when determining each image pixel. In
problems, namely, it is impossible to tell whether a fire is addition, moving lights, such as automobile headlamps, can
occurring until smoke or flame spreads to the detection be mistaken for fires when using only motion information.
range of sensors and detection takes a long time because Therefore, we defined the color conditions of fire features
a fire alarm is only issued after the flame’s influence exceeds in both hue, saturation and value (HSV) and RGB color
space. The local features and motion information of fires
a reference temperature or a set value. In addition, because
were expressed using the mean, variance, and image differ-
the detection range is restricted to the local position where
ence. Moreover, temporally blinking properties were mod-
sensors are installed, the need for numerous distributed sen-
eled by using crossing patches so that it was possible to
sors and installation increases the cost. On the other hand, a
reduce the false detections described above. We defined
fire detection system using image processing does not
22 image features to classify patches into fire or nonfire
require any additional costs, as it uses surveillance cameras areas by using a random-forest-supervised learning algo-
already installed in public places, roads, and tunnels. Thus, rithm. The additive Markov chain was introduced to reduce
fires can be quickly detected without having to wait until false alarms. Also, we designed an embedded surveillance
flames spread to the sensors, and can be monitored over a device made from acrylonitrile butadiene styrene (ABS)
broader range. for stable fire detection in outdoor environments.
Early fire detection algorithms used rule-based color All modules, including image features extraction and fire
models.4–6 They classify the image pixels as fire or not fire classification, are realized to test fire images of different con-
by using the RGB, hue, saturation and intensity, and YCbCr ditions and verified to be robust in complex environments.
color models. However, these methods have many false clas-
sifications. For example, red clothes, moving cars, and lights
will be recognized as fires because only color information 2 Embedded Camera System
was used. Han7 used color and motion information to present In general, to protect people from fires, fire surveillance sys-
real-time fire detection. Töreyin et al.8 used temporal and tems require supervisors in monitoring rooms to observe vid-
spatial wavelet analysis to separate the fire regions from eos transferred from CCTVs. Building a surveillance system
sequential images. Although they presented good experi- requires considerable effort to install surveillance cameras,
mental results, this approach is impractical in various envi- as many repeaters and wiring inside facilities are needed
ronments due to many heuristic thresholds. Z. Teng9 to transfer videos to monitoring rooms, causing increased
modeled the sequentially changing image pixel values by a installation costs. In addition, all images are concentrated
hidden Markov model framework. Ko et al.10 applied a radial on servers in monitoring rooms and so the structure of server
basis function kernel to two-class support vector machines. computers becomes complex. Therefore, we designed an
This research that uses machine learning algorithms could embedded system with a built-in pan-tilt-zoom (PTZ) camera
have an over-fitting problem and are not suitable for various to observe fire areas more effectively. The proposed system,
Optical Engineering 067202-1 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
Fig. 1 Structure of fire detection system. (a) Fire detection system using commercial housing. (b) Customized system showing ABS resin housing
and cooling conductor.
which has a built-in embedded processor inside the camera, connector, as various foreign bodies in the form of particu-
is capable of real-time image processing and consists of a late could otherwise enter the system.
main camera that can monitor fires over a wide range and The whole operation of the fire detection system is
a subcamera that can quickly zoom into fire areas when a depicted in Fig. 2. The main camera is a fixed type and
fire occurs. acquires raw images to detect fire while monitoring the
The early-designed device fixed each module by install- wide surveillance area. In addition, the installed industrial
ing a fixed bracket inside a universal housing, as shown in digital camera is able to remove the noise generated in
Fig. 1(a) and a subcamera below the housing. Due to heat the process of converting images from analog to digital.
from the camera and embedded board, a prolonged operating This can simplify the system and permit control directly
time rapidly raised the internal temperature of the housing. In by the program. A PTZ camera with maximum 10-magnifi-
addition, the commercial housing suffered from the disad- cation was selected for the subcamera so that it is possible to
vantages of heavy weight and expensive cost. In the second immediately magnify and transmit images of the area where
system design, the cooling performance was improved by a fire has occurred. The images acquired from the main cam-
projecting the cooling section of the embedded board out- era are processed in the built-in embedded board, and thus
side, as shown in Fig. 1(b). This system reduced weight fire areas are detected. When the fire areas are detected, mag-
by using an ABS resin housing and a digital output main nified images of fires are acquired from the subcamera. The
camera. The ABS housing has the advantage of reducing acquired images from the subcamera are converted into dig-
production costs because it can be produced in large quan- ital images by a frame grabber. Information, such as a gazing
tities by injection molding. Protection from dust and mois- direction and scale, is transmitted by RS-422 communication
ture was accomplished by using a rubber seal and waterproof to the main processor, for calculating PTZ motion data from
Fig. 2 Operation process of the fire detection system.
Optical Engineering 067202-2 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
Fig. 5 Image Patch by block division of image.
branching for lost data processing and handle both categori-
cal and numerical data. In addition, it makes it unnecessary
to normalize image features and easy to understand the
importance of variables.
First, in order to model the local features of fires, the input
image is divided into small r × c patches (P), as shown in
Fig. 5. We generate the color conditions, motion, and
blink information of fires as features in the patch.
Fig. 3 Magnified images by designating a fire detection area and
using a PTZ camera. 3.1 Modeling Fire Features
In general, fires have specific colors in images. In this sec-
tion, fire features were analyzed in RGB and HSV color
the fire area detection of the main camera. Finally, the system space. For color features of fire in the RGB color space,
transfers the fire location and magnified images of the fire it was observed that the R (red) color is larger than the thresh-
area to the monitoring room. Figure 3 is a graphical user old value RTh and the R color value is the largest in the RGB
interface screenshot of the fire monitoring system and an channel. Moreover, in the HSV color space, it is shown that
image that detects fires and magnifies the detected areas. the H (hue) value of fire is 0 to 70 or 290 to 360. For exam-
Figure 4 shows CPU temperature change in both the cus- ple, the feature f 3 means the saturation will decrease with the
tomized and commercial housing during 24 h. The measure- increasing red component, and thus the expression indicates
ments were conducted after a detection algorithm operated that saturation will decrease downward to zero when the red
for 1 h in each housing. It indicates that average improve- component increases toward the upmost value 255 (Refs. 12,
ment of cooling performance is 22.017°C. 13). Therefore, the patch is likely to indicate a fire area when
there are many pixels satisfying these fire conditions.
3 Fire Detection Algorithm
In previous studies,6–10 information from each pixel in the
image was used to determine whether fires exist in the
images. Alternatively, our paper proposes a method to model
the local characteristics of fire, represented in sequential
images based on patches of an input image, and to detect
a fire area by using a supervised learning algorithm. The ran-
dom forest showed a higher recognition rate than other
machine learning algorithms.11 It can process substitution Fig. 6 Cross extractors.
Table 1 Pseudo-code for the feature calculation of f 1 ∼ f 4 .
Initiailize f 1 ∼ f 4
For all pixels in a patch
ifðPðr ; cÞR > R Th Þ; f 1 þ þ;
ifðPðr ; cÞR > Pðr ; cÞG > Pðr ; cÞB Þ; f 2 þ þ;
ifðPðr ; cÞS ≥ ð255 − Pðr ; cÞR Þ × S Th ∕R Th Þ; f 3 þ þ;
Fig. 4 CPU Temperature changes in customized and commercial ifð290 ≤ Pðr ; cÞH ≤ 360 or 0 ≤ Pðr ; cÞH ≤ 70Þ; f 4 þ þ;
housing during 24 h.
Optical Engineering 067202-3 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
Fig. 7 Feature extraction from a fire image. (a) Original image. (b) Color condition. (c) Motion information. (d) Crossing.
Features (f i ) 1 to 4 show how many pixels satisfy each con-
dition in the patch (Table 1).
The pixels in the fire areas are similar to specific colors
such as red, yellow and white, and the deviation of colors is
relatively uniform. Therefore, Features (f i ) 5 to 10 are
defined by calculating the color mean and the variance for
each RGB channel.
PC ðr; cÞ ∼ NðμC ; σ 2C Þ; C ∈ fR; G; Bg (1)
f ½5: : : 10 ¼ fμR ; μG ; μB ; σ 2R ; σ 2G ; σ 2B g (2)
Additionally, it is observed that fires appear to blink in Fig. 9 Learning and classifying process.
sequential images. Using this, the number of blinks in a cer-
tain area accumulated over 10 frames is defined as a feature.
Fig. 10 Image patches used for learning. Fire sample (a), nonfire
Fig. 8 Random forest. sample (b).
Optical Engineering 067202-4 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
X X
Diff ¼ Iðpi Þ − Iðpi Þ (3) if½Pðc ¼ firejfÞ > Pðc ¼ NonFirejfÞthen Fire else NonFire
pi ∈black area pi ∈white area
(8)
ifjDiff t − Diff t−1 j > threshold; then crossing occurs. (4) The whole process of learning and classification is
described in Fig. 9. In order to learn fire, the input image
We defined three cross extractors to calculate the bright- is divided into the unit area. The 22 features defined in
ness difference between the previous and subsequent patches the previous section are extracted from the divided images.
as shown in Fig. 6. If the difference in brightness exceeds a When learners input the sample groups fc; fg; c ∈
threshold value, then it is considered to be a crossing count. ffire; NonFireg into a learning machine, including classes
Iðpi Þ in Eq. (3) is a brightness value at the pixel pi in the and features generated from the fire and nonfire images as
image. In this paper, a total of nine features were generated shown in Fig. 10, the ensemble classifier and decision
by using three different thresholds deciding the cross state in rule are generated by the random forest. When images are
order to know the level of blinking. captured in real-time, the random forest decides the fire
and nonfire area by the decision rule.
f ½11: : : .19 ¼ fcrossingith1 ; crossingith2 ; crossingith3 g;
(5)
i ∈ f1; 2; 3g.
3.3 Decision of a Fire
Finally, this study used the feature that many motions When using the fire detection algorithm proposed in the pre-
could be observed between previous and present frames in vious section, it is possible to distinguish between fire and
fire areas. Automobile headlamps are known to generate
strong noise, due to the dispersion of light when using
Table 2 Descriptions of experimental movies.
the common image difference between previous and sub-
sequent frames. Therefore, motion pixels are defined as fea-
tures after the removal of the noise by using the average Video sequence Description
brightness of the three previous and subsequent frames.
c t
Movie 1 People set fire to trees using equipment.
Xr X t−1 t−2
f ½20: : : 22 ¼ PC ðr; cÞ þ PC ðr; cÞ þ PC ðr; cÞ
3 Movie 2 There are trees burning at a close distance.
r∈P c∈P
Pt−1 ðr; cÞ þ Pt−2 t−3
C ðr; cÞ þ PC ðr; cÞ
Movie 3 There is a fire in a narrow outside area. This
− C ; movie is used for testing a small fire and a
3 similar color in a scene.
C ∈ fR; G; Bg. (6)
Movie 4 An entire forest is burning with smoke.
Figure 7 is the result that depicts the features defined
Movie 5 There are automobile accidents occurring in
above in an image. The image shown in the upper right indi- the tunnel. This movie is used for confirming
cates the area meeting the color conditions of Features 1 to 4. whether pixels with the similar fire color and
In addition, the image shown in the lower left represents the motion are accurately classified in the scenes.
motion of overlapping frames corresponding to features 20 There is no fire.
to 22. An image shown in the lower right represents the
crossing blocks between the previous and present frames. Movie 6 People create a fire and the equipment
extinguishes the fire inside a building.
3.2 Learning RF and Classifying Fire Features Movie 7 Firefighters set fire to a train in a tunnel.
In order to classify the features defined in the previous sec-
tion into fire and nonfire areas, this study used the random- Movie 8 There is fire occurring in an automobile and
a house in a city.
forest supervised-learning method.14
The input feature (f) reaches the leaf node after following
Movie 9 People set fire in a bucket and there are red
the binary branch of decision tree (tn ). Each node (N) in the objects such as a hat, backpack,
ensembles t1 ðfÞ; t2 ðfÞ; : : : tn ðfÞ selects a feature minimizing and a fire extinguisher.
the Gini impurity (iðNÞ) in order to determine the feature and
threshold value that can classify the feature into a suitable Movie 10 In the movie there is an ambulance blinking a
class most effectively. Finally, probability [PðcjfÞ] of siren. This movie is used for confirming whether
each class is calculated by the summation of each leaf pixels with the similar fire color and motion are
accurately classified in scenes. There is no fire.
node probability, as shown in Fig. 8.
X Movie 11 There are a lot of automobiles reducing speed
iðNÞ ¼ Pðci Þ log Pðcj Þ. (7) at night with headlamps and tail lights on. This
j≠i movie is used for confirming whether pixels
with the similar fire color and motion are
accurately classified in scenes. There is no fire.
The assembly of each ensemble classifier generated in
this manner is called a random forest. When given ensemble NOTE: The movies 1 ∼ 5 are available at http://signal.ee.bilkent.edu
classifiers, feature vector F and class C, the feature vector is .tr/VisiFire/. The movies 9 and 11 were recorded by our embedded
classified into fire as follows: system. The other movies were downloaded on YouTube.
Optical Engineering 067202-5 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
nonfire in various environments. However, false alarms give a higher weight to the fire’s appearance frequency or
occurred in a couple of frames when an automotive head- continuity, and 0.7 is used in this experiment. z is a normal-
lamp and taillight continuously intersected at night. In izing constant.
order to solve this problem, the presence or absence of
fires was finally determined by recording recent fire detec- B ¼ fx1 ; x2 ; : : : ; xN g; xn ∈ f0; 1g (11)
tion results during N frames and stochastically analyzing the
state changes with the use of the additive Markov chain.15
An additive Markov chain is a theory that Xn is affected
by previous sequential random variables (X n−1 ; X n−2 ; : : : )
1 1
and the sum of all the values becomes the final probability. fðxn ; xn−1 Þ ¼ ðαÞ xn þ ð1 − αÞ ðxn & xn−1 Þ ;
2N−n 2N−n
X
N
α ¼ ½0 ∼ 1. (12)
PðXn ¼ xn jX n−1 ; : : : ; X n−N Þ ¼ fðxn ; xn−1 Þ∕z (9)
n¼1
In this study, 2N number of cases is generated. The exe-
cuting speed can be increased by the pre-calculated values
z ¼ PðXn ¼ 0jX n−1 ; : : : ; X n−N Þ þ PðX n ¼ 1jX n−1 ; : : : ; X n−N Þ with the Lookup Table.
The detection results are determined during the accumu-
(10) lated five frames, and the following rule is used for the final
decision:
In this case, a memory function (fðÞ) is defined by the
results of fire detection (B) during N frames as a binary num-
ber. It represents the frequency of the fire detection and con- if PðXn ¼ FirejXn−1 ; : : : ; X n−5 Þ
tinuity during N frames. In addition, the function is > PðXn ¼ NonFirejX n−1 ; : : : ; Xn−5 Þ
generated to give a higher weight to the variable close to
the present state. α is a constant to determine whether to × then Fire else NonFire. (13)
Fig. 11 Movie samples used for comparative experiments. (a) Movie 1. (b) Movie 2. (c) Movie 3. (d) Movie 4. (e) Movie 5.
Optical Engineering 067202-6 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
4 Experiments distributed according to their colors and motions. In selecting
To learn fire features, 10,000 and 40,000 samples of a fire the samples of a nonfire area, learning was performed by
and nonfire, respectively, were extracted from video images, dividing into several areas as follows: areas similar in
such as in a tunnel, downtown, and mountain area. Numeri- color to fires without motion, areas different in color from
cal type feature data were used for the training. The classifier fires with active motion, and areas completely different in
was generated by random-forest supervised-learning. color from fires without motion. In such cases, it is desirable
When using the learning algorithm, the selection of not to select a relatively large number of nonfire samples,
only features but also samples used for learning can have because many errors mistaking nonfire for fire occur
a large impact on the classification results. For example, when the number of fire samples is relatively large.
if most fire samples are red areas, the red areas will be In the first experiment as shown in Fig. 11, Movies 1 to 5
detected well. However, it becomes difficult to detect the yel- were used to compare the fire detection algorithm proposed
low and white areas. When learning only areas with active by Töreyin, Ko, and the method proposed in this paper.
motion, it becomes difficult to detect the areas with the small Additionally, Movies 6 to 11 were used to perform experi-
motion of a flame. Therefore, the fire samples used for learn- ments in more varied environments. The description of the
ing should be selected so that the numbers are evenly images can be found in Table 2. True positive means that a
Fig. 12 Images of fire detection results. (a) Movie 1. (b) Movie 2. (c) Movie 3. (d) Movie 4. (e) Movie 6. (f) Movie 7. (g) Movie 8. (h) Movie 9.
Optical Engineering 067202-7 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
Fig. 13 Result of comparative experiments of Töreyin’s and Ko’s algorithms and detection rate (sum of TP þ TN).
Fig. 14 Movies used for analysis of detection rate. (a) Movie 6. (b) Movie 7. (c) Movie 8. (d) Movie 9. (e) Movie 10. (f) Movie 11.
Optical Engineering 067202-8 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-use
Kim and Kang: Fire detection system using random forest classification for image sequences. . .
fire image is correctly determined as a fire, true negative
means that a nonfire image is correctly determined as a non-
fire, false positive means that a fire image is misjudged as a
nonfire and false negative means a nonfire image is mis-
judged as a fire.
Overall results of the first experiment are shown in
Figs. 12 and 13. The method proposed in this study
shows a higher detection rate than the other two methods.
In this case, the detection rate is the sum of the values of
true positive and true negative. In Movie 1, it was not pos-
sible to consider the motion information properly because
the frame rate of the video was lower than other images.
In Movie 3, the detection rate was lower than other images
because the size of the fire area was smaller than other
patches used in the experiment. Furthermore, in Movie 5,
the false alarm shown in the method by Ko did not occur
in the experiment with the proposed method.
As depicted in Fig. 14, the second experiment was per-
formed by using fires in tunnels, inside buildings and auto-
mobiles, with a blinking ambulance light, continuously
passing automobile headlamps, and tail lights likely to be
mistaken for fires. As shown in Fig. 15, the detection rate
was an average of 96.45% and we could reduce the average
of false negative rate from 2.1% to 1.31% when applying the
additive Markov chain. However, as shown in Fig. 16(a) and
16(b), false positives were shown because the fire area was
small at the moment when the fire was extinguished in Movie
6 and as the fire started in Movie 9. Depicted in Fig. 16(c)
and 16(d), the false negatives occurred because the color,
motion and crossing feature were similar to fires. Movie 9
demonstrated whether the classifier works well when expos-
Fig. 15 Result of detection rate performed in various environments.
(a) Before applying additive Markov chain. (b) After applying additive ing a bag, hat, and fire extinguisher similar in color to flames
Markov chain. in the image. The results showed that it worked well even
when moving the hat and bag. Although there were many
areas that had values similar to the fire features defined
above in Movies 10 and 11, the classifier could work
well because each of the feature values were relatively small.
Fig. 16 Cases of false positive: (a) Movie 6, (b) Movie 9. Cases of false negative: (c) Movie 9, (d) Movie 11.
Optical Engineering 067202-9 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-useKim and Kang: Fire detection system using random forest classification for image sequences. . .
and Technology (Nos 2011-0027213 and 2011-0017228)
and this work was partly supported by the Human
Resources Development program (No. 20114010203080)
of the Korea Institute of Energy Technology Evaluation
and Planning (KETEP) grant funded by the Korea government
Ministry of Trade, Industry and Energy.
References
1. S. Verstockt et al., “Multi-sensor fire detection by fusing visual and non-
visual flame features,” in Image Sig. Process., pp. 333–341, Springer,
Berlin, Heidelberg (2010).
2. S. Bhattacharjee et al., “Wireless sensor network-based fire detection,
alarming, monitoring and prevention system for Bord-and-Pillar coal
Fig. 17 Variable importance of learned random forest. mines,” J. Syst. Softw. 85(3) 571–581 (2012).
3. I. Bosch, S. Gómez, and L. Vergara, “A ground system for early forest
fire detection based on infrared signal processing,” Int. J. Rem. Sens.
It is possible to recognize variable importance in the 32(17) 4857–4870 (2011).
4. T.-H. Chen, P.-H. Wu, and Y.-C. Chiou, “An early fire-detection method
learning process when using random-forest supervised- based on image processing,” in Image Processing, 2004. ICIP'04. 2004
learning. Figure 17 shows that the importance of 22 features International Conference on, Vol. 3, pp. 1707–1710, IEEE (2004).
5. T. Çelik and H. Demirel, “Fire detection in video sequences using a
is evenly distributed. In this case, Features 5 and 7 are rel- generic color model,” Fire Safety J. 44(2) 147–158 (2009).
atively high because the classifier frequently used the mean 6. W.-B. Horng, J.-W. Peng, and C.-Y. Chen, “A new image-based real-
value of red and blue color channels during the training. The time flame detection method using color analysis,” in Networking,
Sensing and Control, 2005. Proceedings. 2005 IEEE, pp. 100–105,
rate of features (1 ∼ 10) related to the fire color are 51.2%. It IEEE (2005).
means that the color features are more effective than the 7. D. Han and B. Lee, “Flame and smoke detection method for early
motion and crossing features (11 ∼ 22). However, the motion real-time detection of a tunnel fire,” Fire Safety J. 44(7) 951–961 (2009).
8. B. Töreyin et al., “Computer vision based method for real-time fire and
and crossing features perform the role of the fire features flame detection,” Patt. Recog. Lett. 27(1) 49–58 (2006).
very well. It works to prevent miss-classifications when sim- 9. Z. Teng, J.-H. Kim, and D.-J. Kang, “Fire detection based on hidden
Markov models,” Intl. J. Contr. Autom. Syst. 8(4) 822–830 (2010).
ilar colors, such as automobile headlamps and red hats, 10. B. C. Ko, K.-H. Cheong, and J.-Y. Nam, “Fire detection based on vision
appear in scenes. In the color features, a contribution of sensor and support vector machines,” Fire Safety J. 44(3), 322–329
the RGB channel values (5 ∼ 7) is higher than the color con- (2009).
11. R. Caruana, N. Karampatziakis, and A. Yessenalina, “An empirical
ditions (1 ∼ 4). In the crossing features (11 ∼ 19), crossing evaluation of supervised learning in high dimensions,” in
features for small threshold (11 ∼ 13) have more influences Proceedings of the 25th international conference on Machine learning,
on detecting fires. pp. 96–103, ACM (2008).
12. T.-H. Chen, P.-H. Wu, and Y.-C. Chiou, “An early fire-detection method
The algorithm could process approximately 30 frames of based on image processing,” in Image Processing, 2004. ICIP'04. 2004
320 × 240 image in Core 2 Duo 2.00 Ghz at our embedded International Conference on, Vol. 3, pp. 1707–1710, IEEE (2004).
13. T.-H. Chen, C.-L. Kao, and S.-M. Chang, “An intelligent real-time fire-
system. detection method based on video processing,” in Security Technology,
2003. Proceedings. IEEE 37th Annual 2003 International Carnahan
5 Conclusions Conference on, pp. 104–111, IEEE (2003).
14. L. Breiman, “Random forests,” Mach. Learn. 45(1) 5–32 (2001).
In this paper, a fire detection system with the built-in 15. S. S. Melnyk, O. V. Usatenko, and V. A. Yampol’Skii, “Memory func-
tions of the additive Markov chains: applications to complex dynamic
embedded processor, the main camera, and the PTZ camera systems,” Phys. Stat. Mech. Appl. 361(2), 405–415 (2006).
was designed to perform fire surveillance in various environ-
ments. In addition, fire features were modeled by color,
Onecue Kim received a BS in computer sci-
motion, and crossing features. This paper proposed an algo- ence from Tongmyong University, Korea, in
rithm that could detect fires precisely by using the random 2008. He is now a postgraduate and doctoral
forest classifier, making it possible to reduce the false alarms student in mechanical engineering at Pusan
that occurred in previous studies resulting from the complex National University, Korea. His current
research interests are visual surveillance,
background. We used an additive Markov chain method for a machine vision, robots, and pattern recogni-
robust fire detection algorithm. The performances of this tion.
method were verified by comparative experiments with
other methods in various environments. The proposed fire
detection system can detect fires with high reliability in pub-
lic places, such as buildings, roads and tunnels. Also, it min- Dong-Joong Kang received a BS in preci-
imizes casualties and property damages by early fire sion engineering from Pusan National Uni-
detection. In addition, the automatic surveillance system versity in 1988 and a PhD in automation
and design engineering at KAIST (Korea
can reduce facility maintenance costs by reducing the num- Advanced Institute of Science and Technol-
ber of monitoring personnel and the fatigue level of the ogy) in 1998. From 1997 to 1999, he was
supervisor. Therefore, our system contributes to increased a research engineer at Samsung Advanced
safety in public places because potential fires can be moni- Institute of Technology (SAIT) and from
2000 to 2005 he was an assistant professor
tored 24 h a day. in the mechatronics engineering department
at Tongmyong University. He has been an
Acknowledgments associate professor at the School of Mechanical Engineering at
Pusan National University since 2006 and associate editor of the
This work was supported by the Basic Science Research International Journal of Control, Automation, and Systems since
Program through the National Research Foundation of 2007. His current research interests include visual surveillance, intel-
Korea (NRF) funded by the Ministry of Education, Science ligent vehicles/robotics, and machine vision.
Optical Engineering 067202-10 June 2013/Vol. 52(6)
Downloaded From: https://www.spiedigitallibrary.org/journals/Optical-Engineering on 22 Nov 2021
Terms of Use: https://www.spiedigitallibrary.org/terms-of-useYou can also read
